At the Finance and AI Research Lab, Queen’s Business School, we recognize the application of artificial intelligence in financial services as a wicked problem:
A problems with combined degrees of conflict (Adelson et al. 2023), complexity (Kelly, Malamud, and Zhou 2022) and uncertainty(Dotan and Ravid 1985).
Financial tail risk, advanced analytics and artificial intelligence Knowledge Transfer Project (UKRI) with Funds-Axis Ltd.
Leveraging Artificial Intelligence to enhance and understand regulatory compliance in the investment management industry with Funds-Axis Ltd.
Towards a trustworthy banking approach to AI implementation: Balancing membership trust with operational performance with Credit Union Development Association.
Figure 1: Relative size of capital markets (2022)
Data sourced from BATS
Figure 2: Economic view of markets
Paradox of the markets
Much like Groucho Mark’s refusal to join a club that would have him as a member,
Everyone should refuse to transact with anyone willing to transact with them!
Fraud, defined as criminal deception for unjust advantage, has evolved with technology (Bolton and Hand 2002).
Fraud detection usually works along side fraud prevention, where there is a necessity for detection methods when prevention fails.
Classic prevention methods include:
Spoofing and closing price manipulation are both forms of market manipulation but they differ in their methods and objectives.
Spoofing is an especially prevasive problem in US stock markets where 97% of orders are cancelled before they trade (Khomyn and Putniņš 2021)
JP Morgan paid over $900Million in fines for spoofing activity in the commodities markets during 2008-2016 (Debie et al. 2023)
- This is an actual FINRA manipulation case from 2016(Zhai, Cao, and Ding 2018)
We have extracted information from the official announcement of administrative penalties.
Which stock was manipulated on which day?
Important
We obtained 43,045 samples, of which 1872 are manipulation cases.
Input: One day time series data for one stock.
Matrix (number of time steps x number of features)
5 minutes sampling, approximately 50 time steps per day.
Output: Normal (0) or abnormal (1).
## Results
## Results
## Next steps